Comment by Frannky | Hacker News Reader

Comment by Frannky | Hacker News Reader

Parent Full thread

I'm using oh my pi with Kimi K3 as the planner and DeepSeek Flash 0731 as the implementer, with OpenRouter as the provider.

Any suggestions for a better configuration? I mostly need Opus 4.6 + Claude Code alternative. I don't need Fable level capabilities, I add one new feature at a time and approve the code before shipping it. Then test and open a PR.

Usually both Fable and Opus suggest dumb ideas but are good implementers once I tweak the idea, approve the code, and add unit and production tests.

I'm OK spending max $100/month on APIs, ideally with Zero data retention. I only need a few hours a day of coding. I don't want agents running all the time; I figure I can stay on top to each feature and wrap my head around the product and new suggestions as long as I don't build too much at once.

I'm still on a Claude Max $100 plan, but it's barely usable anymore—one call and I hit 20–30% of the 5h window on Opus 4.8. Opus 5 seems tuned to make messes, and Fable burns tokens for a level of capability I don't actually need.

> DeepSeek Flash 0731 as the implementer

MiniMax's "token plan" ($20/mo for 1.7b tokens) is cost competitive. MiniMax M3 is equally good, if not better than DeepSeek v4, at coding: https://platform.minimax.io/subscribe/token-plan?tab=individ...

If you prefer pay-as-you-go, then Xiaomi MiMo is the only other provider with comparable models (MiMo v2.5 & Pro) that matches DeepSeek's current API rates for input/output/cache: https://mimo.mi.com/docs/price/pay-as-you-go

Meanwhile, Meta is running a 10x discount on Muse Spark 1.2 (Grok 4.5 / Sonnet 5 level model), if you opt-in to data sharing: https://dev.meta.ai/docs/getting-started/pricing-rate-limits

> I'm OK spending max $100/month on APIs, ideally with Zero data retention.

In that case, probably you'll get more out of OpenAI's coding plan, as (from what I hear routinely) the GPT 5.6 series is thrifty with token use but as smart as the Claude 5 series: https://x.com/ArtificialAnlys/status/2085083490056589784 / https://archive.vn/3VDlN

Token Plan Individual for $6/month has Qwen-3.8-max, HappyHorse1.1, Deepseek V4 and GLM-5.2 models.

https://www.alibabacloud.com/en/campaign/ai-token-plan

not related to your question directly, but noticed you mentioned oh-my-pi, can you share little bit more why you went with oh-my-pi and not install your own set of extensions? (asking because I was just looking at it to enhance my workflow, but feeling it has too many things)

I was in your situation some time ago. Good thing is that advanced (but little bloated, yes) harnesses have good enough defaults, so you don't have to go and read each setting before the use.

Regarding Pi my position is that it is brilliant piece of software if you don't need extensions - give the model bash tool and let it do all it wants through it or use Pi as SDK for your own advanced harness or smth similar.

But Pi with extensions has two problems. First one: rather often they don't play well together. For example, if you want some adjustments A and B for the same tool and there are two extensions which do A and B, they will likely not work as expected when installed simultaneously. You could say that it can be solved by adjusting extensions or just generating your own - yes and it is the second problem. Like any piece of code you own and use, you have to maintain it. Bug here, incompatibility there and voila - you spend your precious time to work on harness instead of doing your job. Plus remember that vibecoders are not very responsible people, so Pi extensions registry is flooded by "use Pi to customize Pi" buggy one shot extensions.

With carefully developed set of extensions Pi would be better, like properly configured Arch Linux could be better than Linux Mint in the hands of power user. But considering how fast things are changing in this sphere, seems it is more optimal to take more bloated harness - with unneeded tools, too big prompts, etc - which will be effective on 90%, but do the actual job with that harness right now.

It's all about trade-offs, do you want a bunch of extensions of varying quality from a bunch of developers, or do you want a bunch of extensions of varying quality from one developer?

Or the third path: use Pi to customize Pi. It’s pretty fast and easy, and now it does exactly what you want.

That's just "a bunch of extensions of varying quality from one developer" with extra steps

I just didn't want to spend time configuring an harness

I'm curious what kind of prompts people are doing to get 'dumb' ideas out of Fable

It's usually one of two things. First, I say I want to build this system solving this problem with these requirements, what do you suggest? It suggests X, then I say, "Hmm, why not Y?" and it goes, "Oh yeah, Y is better," etc.

It's similar to the situation where you ask whether you should walk or take the car to go refill the car with gas. I feel LLMs have a linearity embedded in them that prevents them from finding non-linear, smart solutions, especially when data is scarce. They can probably get it done, but with far more complex solutions, which increases the risk of ending up with a crazy complex codebase when it could have been much simpler and more elegant.

The other situation is when I spot a problem or a feature change that's needed after experiencing the product, and I ask how to change the codebase, it suggests something, I say "why not this other thing," and then we do the other thing.

It’s a tool that does what you want not an opinionated senior dev. Ask it to research and do an evaluation on the pros and cons of choosing X over Y and it will do a fair job.

I have the 20x. I really never reach the limit anymore. I am a heavy user. How is that possible?

I agree with you. I am a 5x user; compared to 6 months ago, I rarely reach session or weekly limits. It got slightly slower, but fewer mistakes make up for it. Probably the commenter above is a heavy vibe coder launching a bunch of background agents to reread the base code over and over.

I'm on 5x also, and lately (this week), the window seems more generous, even on Fable. I wonder what's going on.

Do you import old code base or work on small greenfield projects?

Honestly, GPT 5.6 Luna is worth a look. It’s a reasonably good implementer at a small fraction of the cost. $100 buys a heck of a lot of it at API pricing.

Not sure about the OAi Pro plan, doesn’t look like the 80% Luna price slash made its way into the quota system.

You could also try tuning down the effort level on Opus. It makes a huge difference in token consumption and you might be able to get away with lower than you’ve set

Are you using Kimi w/expectation of ZDR? Their EULA is “we use your data and train on it unless you negotiate something with us privately”

Fireworks.ai has ZDR and hosts Kimi K3

Maybe dumb question: why don't you switch back to 4.6? I am still using it

Reply on news.ycombinator.com